Google unveils Gemini 4 Argon flagship after months of delays, with access limited to cybersecurity partners
Google has finally shipped its delayed flagship. Gemini 4 Argon opens the Gemini 4 generation and is aimed at three areas: software development, cybersecurity and financial operations. According to source reports, Argon can generate up to 1 million output tokens, which allows very long reasoning chains and large code rewrites. It also offers automated vulnerability detection and remediation. Google cancelled Gemini 3.5 Pro to put its resources behind this model.
The rollout is deliberately narrow. Access is limited to verified cybersecurity partners through Google's Fairwind Program, under the US voluntary pre-release review process for frontier models. The New York Times framed it as a flagship released "with limits." Reuters said Argon is positioned directly against OpenAI's Astra and Anthropic's Opus on coding and cybersecurity benchmarks. VentureBeat wrote that Google has retaken the benchmark lead, but only in a limited release. Artificial Analysis has published an independent breakdown of Argon's intelligence, speed and price.
The timing makes Argon part of a crowded week. Anthropic shipped Sonnet 5.5 on Monday. OpenAI shelved GPT-6.1 Astra over safety concerns and launched the cheaper GPT-6.1 Sol at DevDay. Argon may hold the benchmark lead for now, but OpenAI's Astra pause and Google's gated release point to the same trend: labs are attaching formal safety gates to their most capable cyber-relevant models before broad release.
The skeptics have clear points. Community reports say Argon underperforms on real coding tasks despite a 77.9% DeepSWE score. The partner-only release is also drawing "safety theater" accusations from developers who cannot test the model themselves. Watch for three things:
- **General availability timeline:** when developers outside the security program get access.
- **Pricing:** whether the "industry-low price" claim holds once Argon reaches the public API.
- **Independent evaluations:** whether outside tests confirm Google's benchmark parity claims.